| name | memory-qdrant |
| description | Semantic memory recall for AI agents using a local Qdrant vector database. Searches past conversations, matches keywords against a facts file, and routes queries to relevant knowledge files. Use when building agents that need long-term memory, fact verification, or domain-specific knowledge retrieval — all without sending data to the cloud.
|
| license | MIT |
| compatibility | OpenClaw 2026.1.30+, Node.js >= 20, Qdrant (local), mcporter CLI |
| metadata | {"author":"rotomi","version":"2.0.0","platform":"openclaw-plugin"} |
Memory Qdrant — Semantic Recall for AI Agents
Gives your agent a searchable long-term memory backed by a local Qdrant vector database. No cloud, no API keys for storage — your memories stay on your machine.
What It Does
- Semantic Search — finds relevant past conversations using vector similarity (via Qdrant + mcporter)
- Facts Matching — searches a
facts.jsonl file for keyword matches (exact facts that override fuzzy recall)
- Knowledge Routing — maps query keywords to domain-specific knowledge files and hints the agent to read them
- Privacy-First — everything runs locally, no data leaves your machine
Configuration
In your openclaw.json plugins section:
{
"id": "memory-qdrant",
"enabled": true,
"config": {
"serverName": "qdrant-memory",
"factsFile": "/home/user/workspace/memory/facts.jsonl",
"qdrantLimit": 5,
"knowledgeMap": {
"project": "memory/knowledge/projekte.md",
"finance": "memory/knowledge/buchhaltung.md",
"infra": "memory/knowledge/infrastruktur.md"
}
}
}
| Option | Default | Description |
|---|
serverName | qdrant-memory | mcporter server name for Qdrant |
factsFile | — | Path to facts.jsonl (one JSON object per line) |
qdrantLimit | 5 | Max number of semantic search results |
knowledgeMap | {} | Keyword → knowledge file mapping |
Architecture
User message arrives
│
▼
memory-qdrant extracts query from message
│
├──→ Qdrant (via mcporter) → top-N semantic matches
├──→ facts.jsonl → keyword grep → exact facts
└──→ knowledgeMap → keyword match → file hints
│
▼
Agent gets: "ERINNERUNG: [memories + facts + hints]"
Facts File Format
facts.jsonl contains one JSON object per line — human-verified facts that override fuzzy memory:
{"fact": "The API timeout is 30s not 60s — confirmed in production", "date": "2026-03-01", "source": "user"}
{"fact": "We use local Qdrant, not cloud — decided in January", "date": "2026-01-15", "source": "user"}
Knowledge Routing
The knowledgeMap maps keywords in the user's message to knowledge files. When a match is found, the agent receives a hint like:
📚 Relevante Knowledge-Datei: memory/knowledge/buchhaltung.md
This uses progressive disclosure — the agent decides whether to read the full file.
Edge Cases
- Qdrant unavailable — graceful degradation, falls back to facts + knowledge hints only
- Empty query — skipped, no search performed
- No matches — no memory section injected (clean context)
- mcporter timeout — 5s default, logged and skipped
References
